Re: [f2fs-dev] [PATCH] f2fs: get out of a repeat loop when getting a locked data page

[Chao Yu <chao@kernel.org>]

On 2023/4/6 11:18, Jaegeuk Kim wrote:
> On 04/06, Chao Yu wrote:
>> On 2023/4/6 0:39, Jaegeuk Kim wrote:
>>> On 03/27, Matthew Wilcox wrote:
>>>> On Mon, Mar 27, 2023 at 08:30:33AM -0700, Jaegeuk Kim wrote:
>>>>> On 03/26, Chao Yu wrote:
>>>>>> On 2023/3/24 5:39, Jaegeuk Kim wrote:
>>>>>>> https://bugzilla.kernel.org/show_bug.cgi?id=216050
>>>>>>>
>>>>>>> Somehow we're getting a page which has a different mapping.
>>>>>>> Let's avoid the infinite loop.
>>>>>>>
>>>>>>> Cc: <stable@vger.kernel.org>
>>>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
>>>>>>> ---
>>>>>>>     fs/f2fs/data.c | 8 ++------
>>>>>>>     1 file changed, 2 insertions(+), 6 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>>>>> index bf51e6e4eb64..80702c93e885 100644
>>>>>>> --- a/fs/f2fs/data.c
>>>>>>> +++ b/fs/f2fs/data.c
>>>>>>> @@ -1329,18 +1329,14 @@ struct page *f2fs_get_lock_data_page(struct inode *inode, pgoff_t index,
>>>>>>>     {
>>>>>>>     	struct address_space *mapping = inode->i_mapping;
>>>>>>>     	struct page *page;
>>>>>>> -repeat:
>>>>>>> +
>>>>>>>     	page = f2fs_get_read_data_page(inode, index, 0, for_write, NULL);
>>>>>>>     	if (IS_ERR(page))
>>>>>>>     		return page;
>>>>>>>     	/* wait for read completion */
>>>>>>>     	lock_page(page);
>>>>>>> -	if (unlikely(page->mapping != mapping)) {
>>>>>>
>>>>>> How about using such logic only for move_data_page() to limit affect for
>>>>>> other paths?
>>>>>
>>>>> Why move_data_page() only? If this happens, we'll fall into a loop in anywhere?
>>>>>
>>>>>>
>>>>>> Jaegeuk, any thoughts about why mapping is mismatch in between page's one and
>>>>>> inode->i_mapping?
>>>>>
>>>>>>
>>>>>> After several times code review, I didn't get any clue about why f2fs always
>>>>>> get the different mapping in a loop.
>>>>>
>>>>> I couldn't find the path to happen this. So weird. Please check the history in the
>>>>> bug.
>>>>>
>>>>>>
>>>>>> Maybe we can loop MM guys to check whether below folio_file_page() may return
>>>>>> page which has different mapping?
>>>>>
>>>>> Matthew may have some idea on this?
>>>>
>>>> There's a lot of comments in the bug ... hard to come into this one
>>>> cold.
>>>>
>>>> I did notice this one (#119):
>>>> : Interestingly, ref count is 514, which looks suspiciously as a binary
>>>> : flag 1000000010. Is it possible that during 5.17/5.18 implementation
>>>> : of a "pin", somehow binary flag was written to ref count, or something
>>>> : like '1 << ...' happens?
>>>>
>>>> That indicates to me that somehow you've got hold of a THP that is in
>>>> the page cache.  Probably shmem/tmpfs.  That indicate to me a refcount
>>>> problem that looks something like this:
>>>>
>>>> f2fs allocates a page
>>>> f2fs adds the page to the page cache
>>>> f2fs puts the reference to the page without removing it from the
>>>> page cache (how?)
>>>
>>> Is it somewhat related to setting a bit in private field?
>>
>> IIUC, it looks the page reference is added/removed as pair.
>>
>>>
>>> When we migrate the blocks, we do:
>>> 1) get_lock_page()
>>
>> - f2fs_grab_cache_page
>>   - pagecache_get_page
>>    - __filemap_get_folio
>>     - no_page  -> filemap_alloc_folio  page_ref = 1 (referenced by caller)
>>      - filemap_add_folio page_ref = 2 (referenced by radix tree)
>>
>>> 2) submit read
>>> 3) lock_page()
>>> 3) set_page_dirty()
>>> 4) set_page_private_gcing(page)
>>
>> page_ref = 3 (reference by private data)
>>
>>>
>>> --- in fs/f2fs/f2fs.h
>>> 1409 #define PAGE_PRIVATE_SET_FUNC(name, flagname) \
>>> 1410 static inline void set_page_private_##name(struct page *page) \
>>> 1411 { \
>>> 1412         if (!PagePrivate(page)) { \
>>> 1413                 get_page(page); \
>>> 1414                 SetPagePrivate(page); \
>>> 1415                 set_page_private(page, 0); \
>>> 1416         } \
>>> 1417         set_bit(PAGE_PRIVATE_NOT_POINTER, &page_private(page)); \
>>> 1418         set_bit(PAGE_PRIVATE_##flagname, &page_private(page)); \
>>> 1419 }
>>>
>>>
>>> 5) set_page_writebac()
>>> 6) submit write
>>> 7) unlock_page()
>>> 8) put_page(page)
>>
>> page_ref = 2 (ref by caller was removed)
>>
>>>
>>> Later, f2fs_invalidate_folio will do put_page again by:
>>> clear_page_private_gcing(&folio->page);
>>
>> page_ref = 1 (ref by private was removed, and the last left ref is hold by radix tree)
>>
>>>
>>> --- in fs/f2fs/f2fs.h
>>> 1421 #define PAGE_PRIVATE_CLEAR_FUNC(name, flagname) \
>>> 1422 static inline void clear_page_private_##name(struct page *page) \
>>> 1423 { \
>>> 1424         clear_bit(PAGE_PRIVATE_##flagname, &page_private(page)); \
>>> 1425         if (page_private(page) == BIT(PAGE_PRIVATE_NOT_POINTER)) { \
>>> 1426                 set_page_private(page, 0); \
>>> 1427                 if (PagePrivate(page)) { \
>>> 1428                         ClearPagePrivate(page); \
>>
>> Since PagePrivate was cleared, so folio_detach_private in
>> f2fs_invalidate_folio()/f2fs_release_folio will just skip drop reference.
>>
>> static inline void *folio_detach_private(struct folio *folio)
>> {
>> 	void *data = folio_get_private(folio);
>>
>> 	if (!folio_test_private(folio))
>> 		return NULL;
>> 	folio_clear_private(folio);
>> 	folio->private = NULL;
>> 	folio_put(folio);
>>
>> 	return data;
>> }
>>
>> Or am I missing something?
> 
> Ah, I missed folio_test_private() tho, can we really expect get_page(),
> SetPagePrivate(), and set_page_private() is in pair with folio_detach_private()?

I guess we are trying to maintain PagePrivate and page_private w/
inner {set,clear}_page_private_* functions, if they are called in paired correctly,
we don't need to call folio_detach_private() additionally in .release_folio and
.invalid_folio, right? Otherwise there must be a bug.

In this patch, I use bug_on to instead folio_detach_private().
https://lore.kernel.org/linux-f2fs-devel/20230410022418.1843178-1-chao@kernel.org/

In this patch, I use {attach,detach}_page_private() to clean up openned codes.
https://lore.kernel.org/linux-f2fs-devel/20230410022418.1843178-2-chao@kernel.org/

With above two patches, I didn't hit any panic or use-after-free issue when testing
xfstest until now.

Thanks,


> I feel attach/detach_page_private would look better?
> 
>>
>> Thanks,
>>
>>> 1429                         put_page(page); \
>>> 1430                 }\
>>> 1431         } \
>>> 1432 }
>>>
>>>> page is now free, gets reallocated into a THP
>>>> lookup from the f2fs file finds the new THP
>>>> things explode messily
>>>>
>>>> Checking page->mapping is going to avoid the messy explosion, but
>>>> you'll still have a page in the page cache which doesn't actually
>>>> belong to you, and that's going to lead to subtle data corruption.
>>>>
>>>> This should be caught by page_expected_state(), called from
>>>> free_page_is_bad(), called from free_pages_prepare().  Do your testers
>>>> have CONFIG_DEBUG_VM enabled?  That might give you a fighting chance at
>>>> finding the last place which called put_page().  It won't necessarily be
>>>> the _wrong_ place to call put_page() (that may have happened earlier),
>>>> but it may give you a clue.
>>>>
>>>>>>
>>>>>> struct page *pagecache_get_page(struct address_space *mapping, pgoff_t index,
>>>>>> 		int fgp_flags, gfp_t gfp)
>>>>>> {
>>>>>> 	struct folio *folio;
>>>>>>
>>>>>> 	folio = __filemap_get_folio(mapping, index, fgp_flags, gfp);
>>>>>> 	if (IS_ERR(folio))
>>>>>> 		return NULL;
>>>>>> 	return folio_file_page(folio, index);
>>>>>> }
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>>> -		f2fs_put_page(page, 1);
>>>>>>> -		goto repeat;
>>>>>>> -	}
>>>>>>> -	if (unlikely(!PageUptodate(page))) {
>>>>>>> +	if (unlikely(page->mapping != mapping || !PageUptodate(page))) {
>>>>>>>     		f2fs_put_page(page, 1);
>>>>>>>     		return ERR_PTR(-EIO);
>>>>>>>     	}


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel